Your AI can have the best style guide in the world, and still break it without anyone noticing.
That's what happened to us at AutoBoost with the engine that writes our blog and our LinkedIn posts every day. We gave it a written guide with two very specific rules: the word "AI" had to appear within the first words of the headline or the hook, and two posts published the same day couldn't open with the same words. For a while we assumed both were being followed, because they were written down. Then we built an automatic checker and found out that having a rule on paper and having it actually followed are two different things.
A guide is an instruction, not a guarantee
Anyone who puts an AI to work generating content, replies, or documents following a procedure (a brand guide, a customer service script, an accounting classification rule) assumes the same thing: if the instruction is written, it gets followed. It doesn't. A model can drift in one case out of ten, or one out of a hundred, and nobody notices until someone reviews it by hand, which is exactly what automating the process was supposed to avoid.
The fix isn't writing the guide in more detail. It's letting go of trusting the instruction alone and starting to check the output, piece by piece, every day.
Two rules, checked daily, blocking nothing
We added a second layer to the daily alert we already had (which confirms the automated content actually got published): checking that every piece follows two style rules written into the guide.
- The word "AI" within the first words. Matching it as a text fragment isn't enough, because that would count any partial coincidence (a word that happens to contain it) as a pass. The checker compares word by word against the headline or the hook, not by substring.
- Two LinkedIn posts on the same day can't open with the same words. LinkedIn builds each post's URL from the opening words of the text, and the analytics importer matches posts by that URL. If two posts share an opening, the analytics get assigned to the wrong post, and you lose the ability to know which one actually worked.
The checker blocks nothing, it doesn't stop anything from publishing, because the engine that writes runs on its own, with nobody reviewing it live. What it does is alert the same day, over WhatsApp and email, as soon as it catches a violation. That's the difference between finding out today or finding out three weeks from now, with twenty pieces already published carrying the same mistake.
What the real results said
Writing the checker was the easy part. What actually mattered was testing it against something real: does the rule actually predict which pieces perform better?
We reviewed that week's LinkedIn posts against their real performance:
| Post | Follows both rules? | Result |
|---|---|---|
| Post about RAG | Yes | Took off (the best on the account) |
| Post about AI agents | Yes | Took off |
| Rest of that week's posts | No | Fell flat |
| Post about data ("Sort your data before you add AI") | Follows it, but with "AI" in fifth position | Fell flat anyway |
The first two rows confirm the rule: following the guide lines up with the posts that worked. But the last row complicates it: following the letter of the rule (the word was within the opening words) wasn't enough to save a post that had no hook or number inside. The exact position of the word also matters, and that wasn't written down anywhere until the data itself showed it.
That's the part that actually matters to any business putting an AI to work following a procedure: the automatic check can't stay theoretical. You have to test it against real results now and then, because an exception can show up that forces you to refine the rule, not just cases that confirm it.
Why this isn't just a blog thing
The same pattern shows up in any AI working under business rules:
- An AI answering customers over WhatsApp with a fixed script: how do you know if it skipped a required step (asking for a tax ID, confirming the order) in some conversation, without rereading a hundred of them by hand every week?
- An AI classifying invoices by accounting category: how do you know it started misclassifying one specific category, before thirty errors pile up?
- An AI drafting customer communications: how do you know if it dropped a mandatory legal clause from one of them?
In a case we worked on for a veterinary group, the AI that books invoices has this same principle built in by design: it extracts the data, matches the supplier, checks the tax lines add up, and only creates the invoice in the system if the verified total matches to the cent. If it doesn't match, it deletes it. You can see the full case in invoices that book themselves. The underlying lesson is the same as the style checker: giving the AI the right instruction isn't enough, you also have to give it the ability to check its own output before treating it as good.
How to build your own checker
You don't need a complex system to start. The pattern that works for us:
- Write the rule so it can be verified, not left vague. "Sound professional" can't be checked automatically; "have word X within the first N words" can.
- Compare literally, not by intuition. A model checking another model with natural language is cheaper to build, but less reliable than an exact comparison whenever the rule allows one.
- Alert, don't block, at the start. If the process runs without live human oversight, a same-day alert is more realistic than trying to block every mistake with perfect rules from day one.
- Test the rule against real results now and then. A rule nobody checks against reality turns into superstition, whether it's actually being followed or not.
The question you need to ask
If you already have an AI following a procedure in your business (a script, a brand guide, a classification criterion), the question isn't whether you gave it the right instructions. It's whether you'd know, the same day, if it stopped following them.
At AutoBoost we build custom software and AI integrated into your real systems, with the checking layer built in from the design, not bolted on after something goes wrong. You can see how we've built it across different industries in our services and get in touch to talk about your case.
